Problem
I was producing an overview of the medical transcription tools available for use in primary care settings.
Solution
This was one of the areas that really got me interested in the potential of AI and Natural Language Processing. Whisper had just been released by OpenAI and I was excited that I could run this locally and transcribe speech to text for free. No patient identifiable information would be sent over the internet. There were no subscriptions or licences. I found this fascinating and proceeded to make my way through many textbooks of natural language processing.
I made some basic applications in python with Tkinter GUIs. Some of the limiting factors were that the initial models were quite slow on my laptop at the time. There was also limited ability to do live transcription without experimenting with chunking the audio and rejoining the outputs with variable success.
From this I experimented with what has now become 'ambient voice technology'. I extracted the audio from videos on youtube of medical students practicing pretend medical consultations with actors. I could use the transcriptions from the consultations and extract relevant information using an LLM.
Here is my paper from 2023 on NLP in primary care if you would like to have a read. A lot will have moved on and be out of date in this rapidly moving field.